fix(campaign): keep profile matrix evidence exact

[drewstone]

What changed

• Records every distinct paid-call model instead of relabeling mixed calls as the last model.
• Accepts Router MMDD snapshots and rejects routing presets as immutable model identities.
• Rejects duplicate profile identities before any dispatch starts.
• Cancels active sibling profile campaigns after the first failure.
• Includes caller commit, optional dispatch identity, profile identity, and comparison config in resumable cell identity.
• Refuses one profile column that resolves to different snapshots instead of aggregating under a false label.
• Releases TypeScript and Python clients as 0.144.3. Version 0.144.2 was cancelled before either registry published it.

Direct proof

• Focused changed paths: 147/147 passed across 4 files.
• Full suite: 4,753 passed, 3 skipped across 352 files.
• TypeScript and example typechecks passed.
• Biome checked 670 files.
• Build and packed-package verification passed.
• Clean merge tree against fetched origin/main: a2d59145023ee4fe3edf5ed014dd7b0e26785914.

[tangletools]

✅ Auto-approved drewstone PR — b8582b8f

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-08-04T04:25:05Z}

[tangletools]

🟢 Value Audit — sound

Verdict	sound
Concerns	0 (none)
Heuristic	0.0s
Duplication	0.0s
Interrogation	151.6s (2 bridge agents)
Total	151.6s

💰 Value — sound

Hardens profile-matrix model evidence (records every paid-call model, rejects mixed/mismatched snapshots, cancels siblings on failure, adds commit+dispatchRef to cache identity) by correctly extending existing dispatchRef/signal/mapConcurrent primitives — ship.

• What it does: Six integrity tightenings on the profile-matrix path: (1) CampaignCellResult now exposes resolvedModels[] (every distinct receipt model) and only exposes resolvedModel when all receipts agree; (2) run-campaign's modelHasSnapshot is rewritten to accept Router -MMDD snapshots while rejecting @preset/ routing selectors and invalid calendar dates; (3) runProfileMatrix rejects duplicate profile ids pre
• Goals it achieves: Guarantee that every RunRecord's pinned model is the exact model that served every paid call in that cell, so profile comparisons and promotion gates cannot be poisoned by mixed, mislabeled, or silently-relabelled model evidence. This is the integrity follow-up to the concurrent-profile-matrix feature landed in 9b7c5f2 — concurrency made the old 'last receipt wins' relabeling dangerous, and this c
• Assessment: Sound and well-grained. The change correctly reuses three pre-existing primitives rather than inventing new ones: dispatchRef already lived on RunCampaignOptions (run-campaign.ts:55) and is used by benchmarks, fixtures, gepa/skillopt/external-text optimization methods; signal already lived on RunCampaignOptions (run-campaign.ts:49) and was already chained campaignAbort→cellAbort→ctx.signal; mapCon
• Better / existing approach: none — this is the right approach. Searched for existing equivalents: dispatchRef (42 matches across the campaign/contract/benchmarks surfaces) is correctly extended, not reinvented; signal plumbing reuses the existing run-campaign abort chain; no pre-existing duplicate-profile-id validator existed (grep found zero matches before this PR); modelHasSnapshot had exactly two copies (run-record.ts + o
• Model: opencode/zai-coding-plan/glm-5.2
• Bridge attempts: 2
• Bridge warning: opencode/kimi-for-coding/k2p7: opencode: opencode error

🎯 Usefulness — sound

A coherent evidence-integrity hardening of runProfileMatrix that closes real mixed-model relabeling, false-aggregation, and cross-commit cache-reuse holes, threaded through established primitives with no dead surface.

• Integration: Fully wired. runProfileMatrix is the keystone exported from src/campaign/index.ts:259 and consumed by tests/rl/corus-from-matrix.test.ts and the playback adapter. The new resolvedModels field on CampaignCellResult (types.ts:589) is consumed by campaignMeasurementDigest (provenance.ts:580), selfImprove (self-improve.ts:1027), and validated in runProfileMatrix. dispatchRef flows into runCampaign's e
• Fit with existing patterns: Fits the codebase grain exactly. resolvedModels sits alongside the existing resolvedModel on CampaignCellResult and the digest already tracked resolvedModel. dispatchRef mirrors runCampaign.dispatchRef and compareOptimizationMethods.dispatchRef. The signal chain reuses the existing runCampaign→executeCell→ctx.signal wiring rather than inventing a new cancellation path. Duplicate-profile-id and mul
• Real-world viability: Edge paths are covered by new tests and the logic is sound. Mixed models in one cell produce resolvedModels[] with no false resolvedModel (run-campaign.ts:1098-1110); cached-cell resume recomputes evidence from receipts via withCurrentAgentModelEvidence so stale cached models don't leak; moving-alias profiles resolving to different snapshots across cells are rejected; duplicate profile ids rejecte
• Model: opencode/zai-coding-plan/glm-5.2
• Bridge attempts: 1

No concerns — sound change, no better or existing approach found. ✅

---

What this audit checks

It judges the change on its merits — not whether it was tasked out in an issue. Unticketed, fast-moving work is fine; the question is whether the change is good and whether a better or existing approach should be used instead.

Pass	What it asks
Heuristic	Vague title? Whitespace-only or cruft-bearing diff? (content signals only)
Duplication	Do added function/class names already exist elsewhere in the repo?
Value Audit	What does it do? What goal does it achieve? Is it good? Better architecture or already-exists?
Usefulness Audit	Does it integrate and fit? Will it hold up in real use and actually get used?

Findings are concerns, not blocks — the human reviewer decides what to do with them.

_{value-audit · 20260804T042932Z}